/resume-research-hardening

Session handoff from 2026-06-22. Run this at the start of the next session.

---

Step 1 — Verify what landed

Run this first to confirm the two committed tracks are actually in the codebase:

git log --oneline | head -8
# Expect: dcacced feat(research): extend backtest data window to 180d via Alpaca historical API
# Expect: d568fe4 fix(infra): watchdog tz bug + raise orchestrator RAM to 90% + BT divergence analysis

# Confirm data module exists
python -c "from backtesting.data import fetch_bars; print('OK')"

# Confirm lookback flag works
python scripts/backtest_broad_sweep.py --help | grep lookback

If either check fails, something was lost — investigate git log before proceeding.

---

Step 2 — Run inter-day health check

Before writing any code, check that overnight paper trading ran cleanly:

python scripts/system_health_check.py
python -m live_trading.eod_report

Check for CRITICAL errors or missed trading windows (gaps > 15 min during 06:30–13:00 PT in signals JSONL). If the health check shows Overall: ERROR, address that before implementing new features.

---

Step 3 — Implement Track A: Walk-forward validation

Why: Backtest WR was uniformly 68–73% across all 9 strategy classes — that tight clustering signals period-specific parameter fitting on a 60-day window, not real edge. Walk-forward validation splits the data so parameters are never selected on the same bars they are evaluated on.

File: scripts/backtest_broad_sweep.py — grid-search and AC gate section only. Do NOT touch _fetch_one or main() argument parsing (Track 1 owns those — already committed).

What to build:

For every (symbol, strategy, params) combination that passes the existing in-sample AC gate:

1. Split bars: split = int(len(bars) * 0.8) → train_bars, oos_bars
2. Run the existing grid search on train_bars only
3. Take best-params from training, run once on oos_bars
4. OOS must pass: WR ≥ 50% AND expectancy ≥ 0.10%/trade AND n ≥ 5 trades
5. Only both-gate entries enter universe_ac_passes.json
6. Add oos_wr and oos_expectancy fields to each entry; flag oos_skipped: true when OOS window < 200 bars

CLI flag: --no-wfv disables OOS gate (backward compat). WFV is ON by default.

Progress output: [WFV] AAPL RSIOversoldMR: IS WR=74.2% OOS WR=61.3% — PASS

Commit: feat(research): walk-forward validation gate in broad sweep (IS 80% + OOS 20%)

KNOWLEDGE.md: Add Decisions row (2026-06-22) documenting WFV gate and OOS thresholds.

---

Step 4 — Implement Track B: Min-hold exit + bid-ask spread

Why: Win/loss ratios of 0.56–0.62x on RSI/SVWAP strategy classes means exits fire too early on winners. Adding a minimum hold forces the position to ride the mean-reversion longer before a signal exit is allowed. Bid-ask spread not being modeled inflates BT WR by ~3–5%.

Files: backtesting/strategies.py and backtesting/sim_core.py only. Do NOT touch live_trading/ strategy classes.

Part A — Min-hold exit guard

Add min_hold_bars: int = 3 to these classes ONLY (low w/l ratio group):

• RSIOversoldMeanReversion
• SessionVWAPVixMR
• SessionVWAPMeanReversion
• SessionVWAPVolumeMR
• TripleConfirmMeanReversion

Do NOT add to: MeanReversionCombo, BollingerBandsMeanReversion, VWAPMeanReversion, SessionVWAPToDMR (already near-symmetric w/l).

In each exit check:

if bars_held < self.min_hold_bars:
    continue  # skip signal exit

Sweep grid values: [0, 3, 6] — so the optimizer finds optimal minimum hold per symbol.

Add min_hold_bars to scripts/generate_live_registry.py class→param mapping.

Assert min_hold_bars < max_hold for any strategy entry.

Part B — Bid-ask spread cost

In backtesting/sim_core.py PnL calculation, add spread_cost_pct: float = 0.0001:

effective_entry = entry_price * (1 + spread_cost_pct)
effective_exit  = exit_price  * (1 - spread_cost_pct)

Add --spread-cost FLOAT CLI flag to sweep script (default 0.0001). --spread-cost 0 = identical to pre-change behavior.

Tests: pytest tests/ -x -q must pass. Fix failures before committing.

Commit: feat(backtesting): min_hold_bars exit guard + bid-ask spread cost model

KNOWLEDGE.md: Add Decisions row (2026-06-22) for both Part A and Part B.

---

Step 5 — Integration smoke test

After both tracks commit:

# Verify all three switches compose cleanly
python scripts/backtest_broad_sweep.py --symbols AAPL --lookback-days 180 --no-wfv --spread-cost 0
# Should reproduce pre-change behavior

python scripts/backtest_broad_sweep.py --symbols AAPL --lookback-days 180
# Should run full IS/OOS pipeline with Alpaca data

---

Background context

Root cause: Full analysis at docs/PAPER_VS_BACKTEST_ANALYSIS_20260622.md. Short version: 60-day training window + no OOS split + signal-based exits that cut winners early + bid-ask spread not modeled = all nine strategy classes showing 8–35% BT→paper WR gap.

What's already done (do not redo):

• backtesting/data.py — fetch_bars() using Alpaca historical API for 180d bars. Yahoo's 60-day cap is server-side and hard; stitching doesn't work.
• scripts/backtest_broad_sweep.py — --lookback-days N flag, default 180.
• multiframe_runner.py — watchdog timezone fix (no more offset-naive subtract error).
• data/orchestrator_config.json — RAM threshold raised to 90%.

Open design decision still needing user input: EOD force-close (EOD_FORCE_CLOSE=1 in services.ps1). Enabling closes all positions at market close; disabling lets max_hold_cb carry them overnight. Not blocking implementation work — ask user before changing.